platform/upstream/llvm.git
2 years ago[lit] Implement DEFINE and REDEFINE directives
Joel E. Denny [Wed, 21 Sep 2022 15:32:05 +0000 (11:32 -0400)]
[lit] Implement DEFINE and REDEFINE directives

These directives define per-test lit substitutions.  The concept was
discussed at
<https://discourse.llvm.org/t/iterating-lit-run-lines/62596/10>.

For example, the following directives can be inserted into a test file
to define `%{cflags}` and `%{fcflags}` substitutions with empty
initial values, which serve as the parameters of another newly defined
`%{check}` substitution:

```
// DEFINE: %{cflags} =
// DEFINE: %{fcflags} =

// DEFINE: %{check} = %clang_cc1 %{cflags} -emit-llvm -o - %s | \
// DEFINE:            FileCheck %{fcflags} %s
```

The following directives then redefine the parameters before each use
of `%{check}`:

```
// REDEFINE: %{cflags} = -foo
// REDEFINE: %{fcflags} = -check-prefix=FOO
// RUN: %{check}

// REDEFINE: %{cflags} = -bar
// REDEFINE: %{fcflags} = -check-prefix=BAR
// RUN: %{check}
```

Of course, `%{check}` would typically be more elaborate, increasing
the benefit of the reuse.

One issue is that the strings `DEFINE:` and `REDEFINE:` already appear
in 5 tests.  This patch adjusts those tests not to use those strings.
Our prediction is that, in the vast majority of cases, if a test
author mistakenly uses one of those strings for another purpose, the
text appearing after the string will not happen to have the syntax
required for these directives.  Thus, the test author will discover
the mistake immediately when lit reports the syntax error.

This patch also expands the documentation on existing lit substitution
behavior.

Reviewed By: jhenderson, MaskRay, awarzynski

Differential Revision: https://reviews.llvm.org/D132513

2 years ago[NFC] Add GitHub issues to HLSL FIXME comments
Chris Bieneman [Wed, 21 Sep 2022 15:29:26 +0000 (10:29 -0500)]
[NFC] Add GitHub issues to HLSL FIXME comments

In order to make this easier to track I've filed issues for each of the
HLSL FIXME comments that I can find. I may have missed some, but I want
this to be the new default mode.

2 years agoAMDGPU: Move test to correct location
Matt Arsenault [Wed, 21 Sep 2022 15:25:04 +0000 (11:25 -0400)]
AMDGPU: Move test to correct location

This is not a MIR printer/parser test, so it belongs with the ordinary
codegen tests.

2 years ago[mlir][llvm] Support pointer entries in data layout translation
rkayaith [Wed, 7 Sep 2022 16:42:52 +0000 (12:42 -0400)]
[mlir][llvm] Support pointer entries in data layout translation

This adds support for pointer DLTI entries in LLVMIR export, e.g.
```
// translated to: p0:32:64:128
#dlti.dl_entry<!llvm.ptr, dense<[32,64,128]> : vector<3xi32>>
// translated to: p1:32:32:32:64
#dlti.dl_entry<!llvm.ptr<1>, dense<[32,32,32,64]> : vector<4xi32>>
```

Reviewed By: ftynse

Differential Revision: https://reviews.llvm.org/D133434

2 years agoAdd clang_CXXMethod_isDeleted function
Anders Langlands [Wed, 21 Sep 2022 15:12:48 +0000 (11:12 -0400)]
Add clang_CXXMethod_isDeleted function

Adds a function to check if a method has been deleted by copy-pasting
the existing implementation of clang_CXXMethod_isDefaulted and changing
it to call CXXMethod::isDeleted() instead.

Differential Revision: https://reviews.llvm.org/D133924

2 years ago[mlir][memref] Add realloc op.
bixia1 [Wed, 21 Sep 2022 13:59:09 +0000 (06:59 -0700)]
[mlir][memref] Add realloc op.

Add memref.realloc and canonicalization of the op. Add conversion patterns for
lowering the op to LLVM using unaligned alloc or aligned alloc based on the
conversion option.

Add filecheck tests for parsing and converting the op. Add an integration test.

Reviewed By: ftynse

Differential Revision: https://reviews.llvm.org/D133424

2 years ago[SLP]Fix write-after-bounds.
Alexey Bataev [Wed, 21 Sep 2022 14:18:06 +0000 (07:18 -0700)]
[SLP]Fix write-after-bounds.

Mask might be larger than the NumElts-OffsetBeg, need to use actual
indices to avoid acces out of bounds.

2 years ago[InstSimplify] Add additional simplifyWithOpReplaced() vector tests (NFC)
Nikita Popov [Wed, 21 Sep 2022 14:52:25 +0000 (16:52 +0200)]
[InstSimplify] Add additional simplifyWithOpReplaced() vector tests (NFC)

2 years ago[llvm-reduce] Update NoChunks initializer to be in line with AllChunks.
Florian Hahn [Wed, 21 Sep 2022 14:50:29 +0000 (15:50 +0100)]
[llvm-reduce] Update NoChunks initializer to be in line with AllChunks.

Without this patch, the assertion triggers below on the test case,
because we are using different oracles for the verification.

    Assertion failed: (Targets == NoChunksCounter.count() && "number of chunks changes when reducing"), function runDeltaPass, file Delta.cpp, line 272.

2 years ago[NFC][LV] Scalarizing test for masked vector calls
Graham Hunter [Tue, 20 Sep 2022 11:17:32 +0000 (12:17 +0100)]
[NFC][LV] Scalarizing test for masked vector calls

2 years ago[LoopVectorize] Fix test name - the test is for fshl not cttz intrinsic costs
Simon Pilgrim [Wed, 21 Sep 2022 14:24:33 +0000 (15:24 +0100)]
[LoopVectorize] Fix test name - the test is for fshl not cttz intrinsic costs

2 years ago[AIX] change "llvm-nm" to "env OBJECT_MODE=any llvm-nm" in clang/test for AIX OS
zhijian [Wed, 21 Sep 2022 13:38:13 +0000 (09:38 -0400)]
[AIX] change "llvm-nm" to "env OBJECT_MODE=any llvm-nm" in clang/test for AIX OS

Summary:
   since default object mode of llvm-nm is change to -X32 (from default -Xany) in patch https://reviews.llvm.org/D132494.In order not effect the test cases in clang/test we need to change "llvm-nm" to "env OBJECT_MODE=any llvm-nm" in clang/test for AIX OS

Reviewers: James Henderson,David Tenty, Hubert Tong
Differential Revision: https://reviews.llvm.org/D134284

2 years agoLICM: Pass through AssumptionCache
Matt Arsenault [Mon, 19 Sep 2022 20:54:18 +0000 (16:54 -0400)]
LICM: Pass through AssumptionCache

2 years ago[AArch64] add tests for fadd -> fma combines; NFC
Sanjay Patel [Wed, 21 Sep 2022 12:56:25 +0000 (08:56 -0400)]
[AArch64] add tests for fadd -> fma combines; NFC

The transform to create a final fma was added with:
D132837 / c98a46fee6f4043276eac

These tests are intended to show the minimal fast-math-flags
necessary to enable the fold: currently only the final fadd
needs to have "reassoc".

2 years agoAdd all constant physical registers to callee preserved masks
Alex Richardson [Tue, 20 Sep 2022 13:42:32 +0000 (13:42 +0000)]
Add all constant physical registers to callee preserved masks

This allows MachineCopyPropagation to eliminate copies of constant registers
such as zero registers. They were previously not being eliminated as the
check for MO.clobbersPhysReg(AvailSrc) would return true for constant
registers such as MIPS $zero.

To avoid having to manually add the zero registers to all CalleeSavedRegs
instantiations in tablegen, I instead added a new isConstant bit to the
Register and set this for MIPS, RISC-V, and AArch64 zero registers.
RegisterInfoEmitter.cpp looks at this flag and adds all constant registers
to the preserved register mask.

This may also benefit other passes but so far I have only seen differences
in MachineCopyPropagation. In the future it might make sense to generate
`isConstantPhysReg()` from this information.

Original source: https://github.com/CTSRD-CHERI/llvm-project/commit/8588d8b81458ed6d87b674893e7752e6a6915574

Reviewed By: arsenm

Differential Revision: https://reviews.llvm.org/D131958

2 years agoAdd a baseline test for D131958
Alex Richardson [Tue, 20 Sep 2022 13:40:37 +0000 (13:40 +0000)]
Add a baseline test for D131958

This test shows that the save of MIPS $zero to a callee-saved register
is not elided by the machine-cp pass.

Differential Revision: https://reviews.llvm.org/D131957

2 years ago[mlir] Move getDimsOfType to StructuredOpsUtils.h.
Oleg Shyshkov [Wed, 21 Sep 2022 12:47:21 +0000 (12:47 +0000)]
[mlir] Move getDimsOfType to StructuredOpsUtils.h.

Summary:
This change will bring all helpers that work with iterator types to one place.
Currently getDimsOfType is is declared in Linalg.h, but not directly included by
LinalgInterfaces. It worked so far only because all the places that include
LinalgInterfaces.h also include Linalg.h directly or indirectly.

Differential Revision: https://reviews.llvm.org/D134350

2 years ago[CostModel][X86] Add vbmi2 costs for funnelshift/rotate intrinsics
Simon Pilgrim [Wed, 21 Sep 2022 12:48:15 +0000 (13:48 +0100)]
[CostModel][X86] Add vbmi2 costs for funnelshift/rotate intrinsics

Add costs for the funnel shift instructions - fixes some discrepancies I was hitting with costs numbers from the 'cost-tables vs llvm-mca' script D103695

2 years ago[InstCombine] don't canonicalize shl+sub to mul+add
Sanjay Patel [Wed, 21 Sep 2022 12:03:48 +0000 (08:03 -0400)]
[InstCombine] don't canonicalize shl+sub to mul+add

This stops Negator from transforming:
`C1 - shl X, C2 --> mul X, (1<<C2) + C1`
...in the general case. There does not seem to be any analysis
benefit to using mul in IR, and there's definitely downside in
codegen (particularly when the multiply has to be expanded).

If `C1` is 0, then there's a stronger argument that the single
mul is a better canonicalization than negate-of-shl, but we may
want to remove that too.

This was noted as a potential conflict for D133667.

Differential Revision: https://reviews.llvm.org/D134310

2 years agoRevert "[llvm] Remove libcxx, libcxxabi and libunwind from supported LLVM_ENABLE_PROJ...
Louis Dionne [Wed, 21 Sep 2022 12:38:00 +0000 (08:38 -0400)]
Revert "[llvm] Remove libcxx, libcxxabi and libunwind from supported LLVM_ENABLE_PROJECTS"

This reverts commit 887b8bd733ea36aaa5d9b0a2b45df16c6f6bb6a4 while we
work on resolving issues brought up in https://reviews.llvm.org/D132480.

2 years ago[CostModel][X86] Add vbmi2 test coverage for funnelshift/rotate intrinsics
Simon Pilgrim [Wed, 21 Sep 2022 12:34:32 +0000 (13:34 +0100)]
[CostModel][X86] Add vbmi2 test coverage for funnelshift/rotate intrinsics

vbmi2 has vector funnel shift support that we should be costing correctly

2 years ago[CostModel][X86] Remove out of date TODO
Simon Pilgrim [Wed, 21 Sep 2022 11:55:59 +0000 (12:55 +0100)]
[CostModel][X86] Remove out of date TODO

ROTR constant and uniform-constant tests were added some time ago by 2fe1076a0888b4cdc22abc33ef2e5d9a7d8de1b1

2 years ago[clang] AST: SubstTemplateTypeParmType support for non-canonical underlying type
Matheus Izvekov [Sun, 28 Aug 2022 20:26:01 +0000 (22:26 +0200)]
[clang] AST: SubstTemplateTypeParmType support for non-canonical underlying type

This change allows us to represent in the AST some specific
circumstances where we substitute a template parameter type
which is part of the underlying type of a previous substitution.

This presently happens in some circumstances dealing with
substitution of defaulted parameters of template template
parameters, and in some other cases during concepts substitution.

The main motivation for this change is for the future use in the
implementation of template specialization resugaring, as this will
allow us to represent a substitution with sugared types.

Signed-off-by: Matheus Izvekov <mizvekov@gmail.com>
Differential Revision: https://reviews.llvm.org/D132816

2 years ago[mlir] Allow negative strides and offset in StridedLayoutAttr
Ivan Butygin [Sun, 18 Sep 2022 21:34:20 +0000 (23:34 +0200)]
[mlir] Allow negative strides and offset in StridedLayoutAttr

Negative strides are useful for creating reverse-view of array. We don't have specific example for negative offset yet but will add it for consistency.

Differential Revision: https://reviews.llvm.org/D134147

2 years ago[clang] [Driver] Support multiple configuration files
Michał Górny [Tue, 20 Sep 2022 09:04:06 +0000 (11:04 +0200)]
[clang] [Driver] Support multiple configuration files

Support specifying multiple configuration files via multiple `--config`
options.  When multiple files are specified, the options from subsequent
files are appended to the options from the initial file.

While at it, remove the incorrect assertion about CfgFileName being
non-empty.  It can be empty if `--config ""` is passed, and it makes
sense to report it as non-existing file rather than crash.

Differential Revision: https://reviews.llvm.org/D134270

2 years ago[clang][dataflow] Remove deprecated `transfer(const Stmt *, ...)` API.
Wei Yi Tee [Wed, 21 Sep 2022 10:37:52 +0000 (10:37 +0000)]
[clang][dataflow] Remove deprecated `transfer(const Stmt *, ...)` API.

Reviewed By: gribozavr2

Differential Revision: https://reviews.llvm.org/D134318

2 years ago[CostModel][X86] Add CostKinds test coverage for funnelshift/rotate intrinsics
Simon Pilgrim [Wed, 21 Sep 2022 11:00:09 +0000 (12:00 +0100)]
[CostModel][X86] Add CostKinds test coverage for funnelshift/rotate intrinsics

2 years ago[clang][dataflow] Remove deprecated overloads of `checkDataflow` in `TestingSupport.h`.
Wei Yi Tee [Wed, 21 Sep 2022 09:01:00 +0000 (09:01 +0000)]
[clang][dataflow] Remove deprecated overloads of `checkDataflow` in `TestingSupport.h`.

Reviewed By: gribozavr2

Differential Revision: https://reviews.llvm.org/D134081

2 years ago[VE] Remove obsolete ANDrm patterns
Kazushi (Jam) Marukawa [Sun, 18 Sep 2022 06:26:57 +0000 (15:26 +0900)]
[VE] Remove obsolete ANDrm patterns

Remove obsolete ANDrm patterns for MIMM operands.  We add these
translations to optimize commonly used cast operations before
we support MIMM operands directly by each isntruction.  Such
translations are obsolete now.

Reviewed By: efocht

Differential Revision: https://reviews.llvm.org/D134341

2 years ago[InstSimplify] Add vector tests for simplifyWithOpReplaced (NFC)
Nikita Popov [Wed, 21 Sep 2022 10:03:53 +0000 (12:03 +0200)]
[InstSimplify] Add vector tests for simplifyWithOpReplaced (NFC)

2 years ago[ISel] Enable generating more fma instructions.
Thomas Symalla [Fri, 26 Aug 2022 07:14:10 +0000 (09:14 +0200)]
[ISel] Enable generating more fma instructions.

This patch changes a FADD / FMUL => FMA ISel pattern implemented
in D80801 so that it peeks through more than one FMA.

Reviewed By: foad

Differential Revision: https://reviews.llvm.org/D132837

2 years ago[AArch64] Lower scalar sqxtn intrinsics to use fp registers
David Green [Wed, 21 Sep 2022 09:46:43 +0000 (10:46 +0100)]
[AArch64] Lower scalar sqxtn intrinsics to use fp registers

The llvm.aarch64.neon.scalar.sqxtn.i32.i64 intrinsics take and return
integer types, but operate on fp registers. This can create some
inefficiencies in their lowering, where the registers are converted to
fp a little too late. This patch adds lowering for the intrinsics,
creating bitcasts to/from fp types to allow nicer folding later when the
instructions are selected, especially around insert/extracts.

Differential Revision: https://reviews.llvm.org/D134024

2 years ago[AMDGPU] Update checks in mad_u64_u32.ll. NFC.
Jay Foad [Wed, 21 Sep 2022 09:42:23 +0000 (10:42 +0100)]
[AMDGPU] Update checks in mad_u64_u32.ll. NFC.

2 years ago[mlir][tensor][bufferize] Implement getBufferType for Expand/CollapseShapeOp
Matthias Springer [Wed, 21 Sep 2022 02:20:23 +0000 (11:20 +0900)]
[mlir][tensor][bufferize] Implement getBufferType for Expand/CollapseShapeOp

This function must be implemented for all ops, where the result memref type is different from the input memref type.

Differential Revision: https://reviews.llvm.org/D134331

2 years ago[LAA] Fix ICE with scAddExpr in forked pointers
Graham Hunter [Wed, 21 Sep 2022 08:54:14 +0000 (09:54 +0100)]
[LAA] Fix ICE with scAddExpr in forked pointers

The IR from https://github.com/llvm/llvm-project/issues/57368 results
in an assert firing when trying to create a runtime check for the
forked pointer. One of the forks is fine since it's loop invariant,
but the other is a scAddExpr (containing a scAddRecExpr, so not
invariant) when RtCheck::insert expects a scAddRecExpr.

This is a simple fix to just avoid forks which aren't AddRec or
loop invariant. We can allow it as a forked pointer later with
more work.

Reviewed By: fhahn

Differential Revision: https://reviews.llvm.org/D133020

2 years ago[flang] Handle parent component of extended derived-type
Valentin Clement [Wed, 21 Sep 2022 09:19:13 +0000 (11:19 +0200)]
[flang] Handle parent component of extended derived-type

Parent component refers to the parent derived-type of an extended type.
The parent component is skipped when a specififc component is
referred to. This is fine since all the components in extended type
are available in the type itself. When the parent component is referred,
it need to be taken into account correctly.
This patch fixes the case when the parent component is referred. In a
box, an approriate slice is created or updated to point to the first
component of the parent component. For scalar, a simple conversion to
the parent component type is done.

Reviewed By: jeanPerier

Differential Revision: https://reviews.llvm.org/D134170

2 years ago[NFC][Scalarizer] Let testcase be auto-generated from update_test_check
eopXD [Wed, 21 Sep 2022 09:04:27 +0000 (02:04 -0700)]
[NFC][Scalarizer] Let testcase be auto-generated from update_test_check

The current `CHECK` lines are hand-written. Changing them to auto-generated lines.

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D134340

2 years ago[VE][NFC] Change to use l2i/i2l to simplify code
Kazushi (Jam) Marukawa [Sun, 18 Sep 2022 01:26:07 +0000 (10:26 +0900)]
[VE][NFC] Change to use l2i/i2l to simplify code

We previously added l2i/i2l macros to simpily EXTRACT_SUBREG/INSERT_SUBREG
conversions.  This patch changes VEInstrInfo.td to use such macros to
simplify existing code.

Reviewed By: efocht

Differential Revision: https://reviews.llvm.org/D134118

2 years ago[VE] Add maxnum and minnum
Kazushi (Jam) Marukawa [Sat, 17 Sep 2022 09:04:06 +0000 (18:04 +0900)]
[VE] Add maxnum and minnum

Add maxnum and minnum for float and double.  Lowering is already
implemented, so this patch changes them legal and adds regression
tests.

Reviewed By: efocht

Differential Revision: https://reviews.llvm.org/D134108

2 years ago[VE] Change to expand FMA
Kazushi (Jam) Marukawa [Sat, 17 Sep 2022 09:04:06 +0000 (18:04 +0900)]
[VE] Change to expand FMA

VE has fused multiply-add instruction for only vector calculations.  This
patch forces to expand scalar FMA to multiply and add instructions.
This patch also adds regression test.

Reviewed By: efocht

Differential Revision: https://reviews.llvm.org/D134107

2 years ago[AArch64] Insert/Extract of bitcast patterns
David Green [Wed, 21 Sep 2022 08:54:17 +0000 (09:54 +0100)]
[AArch64] Insert/Extract of bitcast patterns

This adds some quick tablegen patterns for vector_insert(bitcast(..))
and bitcast(vector_extract(..)), allowing us to avoid a round-trip
through GPRs.

Differential Revision: https://reviews.llvm.org/D134022

2 years ago[AArch64][SME] Disable inlining when SME attributes require smstart/smstop or lazy...
David Sherwood [Tue, 20 Sep 2022 15:28:34 +0000 (16:28 +0100)]
[AArch64][SME] Disable inlining when SME attributes require smstart/smstop or lazy-save.

Inlining must be disabled when the call-site needs to toggle PSTATE.SM or
when the callee's function body is executed in a different streaming mode than
its caller. This is needed because function calls are the boundaries for
streaming mode changes.

More details about the SME attributes and design can be found
in D131562.

Differential Revision: https://reviews.llvm.org/D131581

2 years ago[MemorySSA] Remove PerformedPhiTranslation flag
Nikita Popov [Wed, 7 Sep 2022 08:20:20 +0000 (10:20 +0200)]
[MemorySSA] Remove PerformedPhiTranslation flag

I believe this is no longer necessary, as the underlying problem
has been fixed in a different way: Nowadays, we will adjust the
location size to beforeOrAfterPointer() if the pointer is not loop
invariant. This makes merging results translated across loop
backedges safe.

The two tests in phi-translation.ll show an improvement while still
being correct: The loads in the loop no longer alias with noalias
pointers, but still alias with the store in the entry block (which
they originally did not -- this is the bug that
PerformedPhiTranslation originally fixed).

Differential Revision: https://reviews.llvm.org/D133404

2 years ago[flang] Limit shape inquiries rewrite to associate construct entity
Jean Perier [Wed, 21 Sep 2022 08:23:45 +0000 (10:23 +0200)]
[flang] Limit shape inquiries rewrite to associate construct entity

The previous code was rewriting all shape inquires on associate
construct entities to inquires on the associated expression or variable.

This is is incorrect because at the point of inquiry, some statement
between the association and the inquiry may have modified the expression
operands or variable in a way that changes its shapes or bounds.

For instance, in the example below, expression rewrites was previously
replacing `size(x, 1)` by `size(p, 1)` which is invalid if p is a
pointer.

```
associate(x => p + 1)
 call call_that_may_modify_p_shape()
 print *, size(x, 1)
end associate
```

This change restricts rewrites of shape inquiries on associate construct entity
to use the associated expression shape and bounds if and only if the
shape/bounds are compile time constant. Otherwise, this may be invalid.

Differential Revision: https://reviews.llvm.org/D133857

2 years ago[NFC][AMDGPU] Pre-commit test for D132837.
Thomas Symalla [Wed, 21 Sep 2022 08:15:55 +0000 (10:15 +0200)]
[NFC][AMDGPU] Pre-commit test for D132837.

Pre-commit an additional fmac test for D132837.

2 years ago[SROA] Check typeSizeEqualsStoreSize in isVectorPromotionViable
Bjorn Pettersson [Fri, 16 Sep 2022 12:54:40 +0000 (14:54 +0200)]
[SROA] Check typeSizeEqualsStoreSize in isVectorPromotionViable

Commit de3445e0ef15c4209 (https://reviews.llvm.org/D132096) made
changes to isVectorPromotionViable basically doing

  // Create Vector with size of V, and each element of type Ty
  ...
  uint64_t ElementSize = DL.getTypeStoreSizeInBits(Ty).getFixedSize();
  uint64_t VectorSize = DL.getTypeSizeInBits(V).getFixedSize();
  ...
  VectorType *VTy = VectorType::get(Ty, VectorSize / ElementSize, false);

Not quite sure why it uses the TypeStoreSize for the ElementSize,
but the new vector would only match in size with the old vector in
situations when the TypeStoreSize equals the TypeSize for Ty.
Therefore this patch adds a typeSizeEqualsStoreSize check as yet
another condition for allowing the the new type as a promotion
candidate.

Without this fix the new @test15 test would fail with an assert
like this:

opt: ../lib/Transforms/Scalar/SROA.cpp:1966:
  auto isVectorPromotionViable(llvm::sroa::Partition &,
                               const llvm::DataLayout &)
       ::(anonymous class)::operator()(llvm::VectorType *,
                                       llvm::VectorType *) const:
  Assertion `DL.getTypeSizeInBits(RHSTy).getFixedSize() ==
             DL.getTypeSizeInBits(LHSTy).getFixedSize() &&
             "Cannot have vector types of different sizes!"' failed.
 ...
 #8  isVectorPromotionViable(...)::$_10::operator()...
 #9  llvm::SROAPass::rewritePartition(...)
#10  llvm::SROAPass::splitAlloca(...)
#11  llvm::SROAPass::runOnAlloca(...)
#12  llvm::SROAPass::runImpl(...)
#13  llvm::SROAPass::run(...)

Reviewed By: MatzeB

Differential Revision: https://reviews.llvm.org/D134032

2 years ago[mlir] Update tests to use kEmitAccessorPrefix_Prefixed
River Riddle [Wed, 21 Sep 2022 01:11:08 +0000 (18:11 -0700)]
[mlir] Update tests to use kEmitAccessorPrefix_Prefixed

These aren't user facing, so just flip them directly to prefixed.

2 years ago[mlir][NFC] Tidy up the doc comment on CastInfo for Types
River Riddle [Wed, 21 Sep 2022 01:12:32 +0000 (18:12 -0700)]
[mlir][NFC] Tidy up the doc comment on CastInfo for Types

This had quite a few grammatical errors and typos.

2 years ago[mlir] Update Attributes to use the new casting infra
River Riddle [Tue, 20 Sep 2022 23:42:59 +0000 (16:42 -0700)]
[mlir] Update Attributes to use the new casting infra

This allows for using the llvm namespace cast methods instead
of the ones on the Attribute class. The Attribute class methods
are kept for now, but we'll want to remove these eventually
(with a really long lead time).

Differential Revision: https://reviews.llvm.org/D134327

2 years ago[mlir][Bazel] Port f5fe92f69385
Adrian Kuegel [Wed, 21 Sep 2022 05:51:36 +0000 (07:51 +0200)]
[mlir][Bazel] Port f5fe92f69385

2 years ago[libc++] Use Clang-16 for c++17.
Mark de Wever [Sat, 27 Aug 2022 11:43:51 +0000 (13:43 +0200)]
[libc++] Use Clang-16 for c++17.

Reviewed By: ldionne, #libc

Differential Revision: https://reviews.llvm.org/D132788

2 years ago[MachineCycle][NFC] add a cache for block and its top level cycle
Chen Zheng [Fri, 16 Sep 2022 05:48:44 +0000 (01:48 -0400)]
[MachineCycle][NFC] add a cache for block and its top level cycle

This solves https://github.com/llvm/llvm-project/issues/57664

Reviewed By: sameerds

Differential Revision: https://reviews.llvm.org/D134019

2 years ago[clang] [docs] Improve formatting & fix typo in config docs
Michał Górny [Tue, 20 Sep 2022 09:29:19 +0000 (11:29 +0200)]
[clang] [docs] Improve formatting & fix typo in config docs

Fix teletype formatting in configuration file documentation to use
double backticks rather than single backticks.  Cover some more names
with this formatting.  Correct the name of config file for `clang-cl`
invocation.

Differential Revision: https://reviews.llvm.org/D134271

2 years ago[Clang][OpenMP] Codegen generation for has_device_addr claues.
Jennifer Yu [Tue, 20 Sep 2022 08:38:34 +0000 (01:38 -0700)]
[Clang][OpenMP] Codegen generation for has_device_addr claues.

This patch add codegen support for the has_device_addr clause. It use
the same logic of is_device_ptr. But passing &var instead pointer to var
to kernal.

Differential Revision: https://reviews.llvm.org/D134268

2 years agoRevert "[mlir][tensor] Support more cases in MergeConsecutiveExtractSlice"
Mehdi Amini [Wed, 21 Sep 2022 03:54:32 +0000 (03:54 +0000)]
Revert "[mlir][tensor] Support more cases in MergeConsecutiveExtractSlice"

This reverts commit 5d4603a02d0c3e0106b10d245322b1d2072c0c3d.

The Dialect/Tensor/fold-consecutive-insert-extract-slice.mlir test is
failing when built with GCC

2 years agoFix -fuse-ld to be linker flag in feature detection check
Kevin Gleason [Tue, 20 Sep 2022 21:49:08 +0000 (21:49 +0000)]
Fix -fuse-ld to be linker flag in feature detection check

Discovered an issue working in StableHLO when attempting to build with `-Wall -Werror`:
https://github.com/openxla/stablehlo/pull/137

Currently, if `LLVM_USE_LLD` and `-DCMAKE_CXX_FLAGS="${CMAKE_CXX_FLAGS} -Wall -Werror"`
are both specified for build, the build will error with:

```
-- Performing Test CXX_SUPPORTS_CUSTOM_LINKER - Failed
CMake Error at /usr/local/google/home/gleasonk/Coding/llvm-build/lib/cmake/llvm/HandleLLVMOptions.cmake:309 (message):
  Host compiler does not support '-fuse-ld=lld'
...

$ cat <build_dir>/CMakeFiles/CMakeError.log
...
clang: error: argument unused during compilation: '-fuse-ld=lld' [-Werror,-Wunused-command-line-argument]
```

It looks like other repos have hit this same issue:
- https://github.com/golang/go/issues/41527 (mentioned in comment)
- https://github.com/iree-org/iree/pull/7450

This can be reproduced in llvm-project with the following build command:

```
# Compile command taken from https://mlir.llvm.org/getting_started/
# and modified for use case
cmake -G Ninja ../llvm \
   -DLLVM_ENABLE_PROJECTS=mlir \
   -DLLVM_BUILD_EXAMPLES=ON \
   -DLLVM_TARGETS_TO_BUILD="X86;NVPTX;AMDGPU" \
   -DCMAKE_BUILD_TYPE=Release \
   -DLLVM_ENABLE_ASSERTIONS=ON \
   -DCMAKE_CXX_COMPILER=clang++ \
   -DCMAKE_C_COMPILER=clang \
   -DLLVM_ENABLE_LLD=ON \
   -DCMAKE_CXX_FLAGS="${CMAKE_CXX_FLAGS} -Wall -Werror"
```

Reviewed By: mehdi_amini

Differential Revision: https://reviews.llvm.org/D134206

2 years ago[mlir][SCF] Fix loop pipelining unable to handle ops with regions
Christopher Bate [Tue, 20 Sep 2022 20:54:35 +0000 (14:54 -0600)]
[mlir][SCF] Fix loop pipelining unable to handle ops with regions

This change allows the SCF LoopPipelining transform to handle ops with
nested regions within the pipelined `scf.for` body. The op and nested
regions are treated as a single unit from the transform's perspective.
This change also makes explicit the requirement that only ops whose
parent Block is the loop body Block are allowed to be scheduled by the
caller.

Reviewed By: ThomasRaoux, nicolasvasilache

Differential Revision: https://reviews.llvm.org/D133965

2 years ago[LLDB][RISCV] Fix UndefinedBehaviorSanitizer in RISCVEmulatorTester
Emmmer [Wed, 21 Sep 2022 03:52:04 +0000 (11:52 +0800)]
[LLDB][RISCV] Fix UndefinedBehaviorSanitizer in RISCVEmulatorTester

2 years ago[RISCV] Remove support for the unratified Zbt extension.
Craig Topper [Wed, 21 Sep 2022 03:26:48 +0000 (20:26 -0700)]
[RISCV] Remove support for the unratified Zbt extension.

This extension does not appear to be on its way to ratification.

Out of the unratified bitmanip extensions, this one had the
largest impact on the compiler.

Posting this patch to start a discussion about whether we should
remove these extensions. We'll talk more at the RISC-V sync meeting this
Thursday.

Reviewed By: asb, reames

Differential Revision: https://reviews.llvm.org/D133834

2 years ago[Driver] Fix -f[no-]unwind-tables -Wunused-command-line-argument after 4388b56d525c08...
Fangrui Song [Wed, 21 Sep 2022 03:13:30 +0000 (20:13 -0700)]
[Driver] Fix -f[no-]unwind-tables -Wunused-command-line-argument after 4388b56d525c08ce3cf941cfbad2428b0e1695b0

2 years ago[RISCV] Support peephole optimization to fold vmerge.vvm that has tail agnostic polic...
jacquesguan [Tue, 30 Aug 2022 09:17:55 +0000 (17:17 +0800)]
[RISCV] Support peephole optimization to fold vmerge.vvm that has tail agnostic policy and unmasked intrinsics.

This patch supports the tail agnostic part of D130442.

Reviewed By: fakepaper56

Differential Revision: https://reviews.llvm.org/D132923

2 years ago[PowerPC][NFC] Add virtual call to show redundant spill of vector registers
Ting Wang [Wed, 21 Sep 2022 01:21:06 +0000 (21:21 -0400)]
[PowerPC][NFC] Add virtual call to show redundant spill of vector registers

Reviewed By: lkail

Differential Revision: https://reviews.llvm.org/D133921

2 years ago[modules] Fix error "malformed or corrupted AST file: 'SourceLocation remap refers...
Volodymyr Sapsai [Tue, 20 Sep 2022 00:46:01 +0000 (17:46 -0700)]
[modules] Fix error "malformed or corrupted AST file: 'SourceLocation remap refers to unknown module...'".

When a framework can be found at a new location, all references to it in
the module cache become outdated. When we try to load such outdated .pcm
file, we shouldn't change any already loaded and processed modules.

If `Module` has `ASTFile`, it means we've read its AST block already and
it is too late to undo that. If `ASTFile` is `None`, there is no value
in setting it to `None` again. So we don't reset `ASTFile` in
`ModuleManager::removeModules` at all.

rdar://97216258

Differential Revision: https://reviews.llvm.org/D134249

2 years ago[clang] Fix an unused variable warning
Kazu Hirata [Wed, 21 Sep 2022 00:41:58 +0000 (17:41 -0700)]
[clang] Fix an unused variable warning

This patch fixes:

  clang/lib/Driver/ToolChains/PS4CPU.cpp:159:14: error: unused
  variable 'IsPS5' [-Werror,-Wunused-variable]

2 years ago AMDGPU: Implicit kernel arguments related optimization when uniform-workgroup-size...
Changpeng Fang [Wed, 21 Sep 2022 00:25:52 +0000 (17:25 -0700)]
 AMDGPU: Implicit kernel arguments related optimization when uniform-workgroup-size=true

 Summary:
   Under code object version 5, ockl_get_local_size returns the value computed by the expression:
workgroup_id < hidden_block_count ? hidden_group_size : hidden_remainder
For functions with the attribute uniform-work-group-size=true. we can evaluate workgroup_id < hidden_block_count
as true, and thus hidden_group_size is returned for ockl_get_local_size.
  With uniform-workgroup-size=true, this work also set all remainders to zero, and if there
is reqd_work_group_size, we also set work-group-size to the required value from the metadata.

Reviewers:
  arsenm and bcahoon

Differential Revision:
  https://reviews.llvm.org/D131276

2 years ago[mlir][tensor] NFC: name various Transforms/ files consistently
Lei Zhang [Wed, 21 Sep 2022 00:16:42 +0000 (20:16 -0400)]
[mlir][tensor] NFC: name various Transforms/ files consistently

Use a suffix to make clear what the contents inside each file
are.

Reviewed By: ThomasRaoux

Differential Revision: https://reviews.llvm.org/D134202

2 years ago[mlir][tensor] Support more cases in MergeConsecutiveExtractSlice
Lei Zhang [Wed, 21 Sep 2022 00:00:49 +0000 (20:00 -0400)]
[mlir][tensor] Support more cases in MergeConsecutiveExtractSlice

This commit adds utility functions to perform general merging of
OffsetSizeAndStrideOpInterface by supporting producer rank
reducing and non-unit strides.

With it we can extend MergeConsecutiveExtractSlice to support
more cases.

Co-authored-by: Mahesh Ravishankar <ravishankarm@google.com>
Reviewed By: ThomasRaoux

Differential Revision: https://reviews.llvm.org/D134294

2 years ago[ELF] Make -V an alias for -v
Fangrui Song [Wed, 21 Sep 2022 00:12:44 +0000 (17:12 -0700)]
[ELF] Make -V an alias for -v

In GNU ld,

* --version skips linker input processing.
* -v and -V keep processing if there is any input file. -V has more
  information we don't support.

We currently make -V an alias for --version which skips input processing.
On many `*-freebsd` and `powerpc-*` targets, `gcc -v` passes `-V` to ld
and expects to process input. Make -V an alias for -v to provide
compatibility.

Fix https://github.com/llvm/llvm-project/issues/57859

2 years ago[ELF][test] Improve -v and --version tests
Fangrui Song [Wed, 21 Sep 2022 00:05:35 +0000 (17:05 -0700)]
[ELF][test] Improve -v and --version tests

2 years agoChange isLittleEndian to follow llvm style and add an accessor
Shubham Sandeep Rastogi [Tue, 20 Sep 2022 23:58:14 +0000 (16:58 -0700)]
Change isLittleEndian to follow llvm style and add an accessor

Differential Revision: https://reviews.llvm.org/D134290

2 years ago[mlir][tensor] Merge consecutive insert_slice/extract_slice ops
Lei Zhang [Tue, 20 Sep 2022 23:52:19 +0000 (19:52 -0400)]
[mlir][tensor] Merge consecutive insert_slice/extract_slice ops

Consecutive tensor.insert_slice/tensor.extract_slice can be
created for the case like tiling convolution and then downsizing
2-D convolutions into 1-D ones. It hinders further transformations.
So adding these patterns to clean it up.

Given that bufferization is sensitive and have requirements over
the IR structure (see https://reviews.llvm.org/D132666),
these patterns are put in Transforms/ with separate entry points
for explicit collection.

Reviewed By: ThomasRaoux, mravishankar

Differential Revision: https://reviews.llvm.org/D133871

2 years ago[libc] add strerror
Michael Jones [Fri, 16 Sep 2022 18:19:38 +0000 (11:19 -0700)]
[libc] add strerror

Strerror maps error numbers to strings. Additionally, a utility for
mapping errors to strings was added so that it could be reused for
perror and similar.

Reviewed By: sivachandra

Differential Revision: https://reviews.llvm.org/D134074

2 years ago[mlir][arith] Fix constant naming in integration tests. NFC.
Jakub Kuderski [Tue, 20 Sep 2022 23:00:32 +0000 (19:00 -0400)]
[mlir][arith] Fix constant naming in integration tests. NFC.

Suggested by @antiagainst in D134321.

2 years ago[mlir][arith] Add integration tests for addi emulation
Jakub Kuderski [Tue, 20 Sep 2022 22:51:20 +0000 (18:51 -0400)]
[mlir][arith] Add integration tests for addi emulation

This includes tests with the exact expected values and comparison-based tests.

Reviewed By: antiagainst

Differential Revision: https://reviews.llvm.org/D134321

2 years ago[DSE] Add value type info checks for masked store candidates in Dead Store Elimination.
Michael Berg [Tue, 20 Sep 2022 22:54:16 +0000 (15:54 -0700)]
[DSE] Add value type info checks for masked store candidates in Dead Store Elimination.

The type information of the store values can diverge when checking for valid
mask store candidates to eliminate via DSE. This patch checks for equivalence
wrt to size and element count.

Reviewed By: fhahn, rui.zhang

Differential Revision: https://reviews.llvm.org/D132700

2 years ago[clang] Fix missing template arguments in AST of access to member variable template
Matheus Izvekov [Tue, 20 Sep 2022 13:18:20 +0000 (15:18 +0200)]
[clang] Fix missing template arguments in AST of access to member variable template

Signed-off-by: Matheus Izvekov <mizvekov@gmail.com>
Differential Revision: https://reviews.llvm.org/D134295

2 years ago[mlir][spirv] Query target environment for mapping memory space
Stanley Winata [Tue, 20 Sep 2022 22:28:40 +0000 (15:28 -0700)]
[mlir][spirv] Query target environment for mapping memory space

Checks spirv::TargetEnv from op to see if it contains either Kernel or Shader capabilities.
If it does, then it will set the memory space mapping accordingly.

Reviewed By: antiagainst

Differential Revision: https://reviews.llvm.org/D134317

2 years ago[HLSL] Support PCH for cc1 mode
Xiang Li [Mon, 22 Aug 2022 23:57:48 +0000 (16:57 -0700)]
[HLSL] Support PCH for cc1 mode

Add HLSLExternalSemaSource as ExternalSemaSource instead of ASTContext::ExternalSource when PCH is included.

This allows a different external source to be set for the AST context.

Reviewed By: beanz

Differential Revision: https://reviews.llvm.org/D132421

2 years ago[mlir][spirv] Add casting ops to/from generic storage space
Nirvedh Meshram [Mon, 19 Sep 2022 18:05:58 +0000 (11:05 -0700)]
[mlir][spirv] Add casting ops to/from generic storage space

Reviwed By: antiagainst

Differential Revision: https://reviews.llvm.org/D134217

2 years ago[PS4] Always enable the .debug_aranges section when using LTO
Matthew Voss [Tue, 20 Sep 2022 21:44:56 +0000 (14:44 -0700)]
[PS4] Always enable the .debug_aranges section when using LTO

This flag enables the .debug_aranges section by passing a flag to LLD
and our internal linker. This also adds a new routine that will generate
the correct flag for our internal linker or set of flags for LLD when
given a list of LLVM options. That ensures multiple LLVM codegen options
can be passed to either linker consistently.

Differential Revision: https://reviews.llvm.org/D134296

2 years ago[LLD][COFF] Support /MAPINFO flag
Pengxuan Zheng [Tue, 20 Sep 2022 00:21:21 +0000 (17:21 -0700)]
[LLD][COFF] Support /MAPINFO flag

This patch adds support for link.exe's /MAPINFO flag to lld-link.exe.

Here is a description of the flag from Microsoft
(https://learn.microsoft.com/en-us/cpp/build/reference/mapinfo-include-information-in-mapfile?view=msvc-170):

 The /MAPINFO option tells the linker to include the specified information in a
 mapfile, which is created if you specify the /MAP option. EXPORTS tells the
 linker to include exported functions.

Reviewed By: rnk

Differential Revision: https://reviews.llvm.org/D134247

2 years ago[mlir][linalg] Swap tensor.extract_slice(linalg.fill)
Lei Zhang [Tue, 20 Sep 2022 21:31:16 +0000 (17:31 -0400)]
[mlir][linalg] Swap tensor.extract_slice(linalg.fill)

This commit adds a pattern to swap

```
tensor.extract_slice(linalg.fill(%cst, %init))
```
into
```
linalg.fill(%cst, tensor.extract_slice(%init))
```
when the linalg.fill op have no other users.
This helps to reduce the fill footprint.

Reviewed By: mravishankar

Differential Revision: https://reviews.llvm.org/D134102

2 years ago[mlir][spirv] Lower max/min vector.reduction for OpenCL
Stanley Winata [Tue, 20 Sep 2022 21:16:52 +0000 (17:16 -0400)]
[mlir][spirv] Lower max/min vector.reduction for OpenCL

Templatizing vector reduction to enable lowering from
vector.reduction max/min to CL ops.

Reviewed By: antiagainst

Differential Revision: https://reviews.llvm.org/D134313

2 years ago[NFC] Fix typo in comment
Louis Dionne [Tue, 20 Sep 2022 21:22:00 +0000 (17:22 -0400)]
[NFC] Fix typo in comment

2 years ago[flang] Add semantics test for atomic_add subroutine
Naje George [Tue, 9 Aug 2022 00:07:42 +0000 (17:07 -0700)]
[flang] Add semantics test for atomic_add subroutine

Reviewed By: ktras

Differential Revision: https://reviews.llvm.org/D131535

2 years ago[HLSL] remove unnecessary abs attributes
Joshua Batista [Tue, 20 Sep 2022 21:04:08 +0000 (16:04 -0500)]
[HLSL] remove unnecessary abs attributes

remove abs non-elementwise attribute statements, stick to elementwise.

Reviewed By: beanz

Differential Revision: https://reviews.llvm.org/D134312

2 years ago[GISel] TreeMatcher: always skip leaves if they don't care
Kai Nacke [Mon, 19 Sep 2022 16:03:32 +0000 (16:03 +0000)]
[GISel] TreeMatcher: always skip leaves if they don't care

In `GIMatchTreeOpcodePartitioner::applyForPartition()`, the loop over
the possible leaves skip a leaf if the instruction does not care
about the instruction.
When processing the referenced operands in the next loop the same
leaves need to be skipped.

Later, when these leaves are added to all partitions, the bit vector
must be resized first before the bit representing the leaf is set.

This fixes a crash in llvm-tblgen.

Reviewed By: arsenm

Differential Revision: https://reviews.llvm.org/D134192

2 years agoAtomicExpand: Use correct pointer size for integer
Matt Arsenault [Tue, 20 Sep 2022 18:58:52 +0000 (14:58 -0400)]
AtomicExpand: Use correct pointer size for integer

This was using the default address space.

2 years agoAMDGPU: Add baseline test for expansion of 16-bit local atomics
Matt Arsenault [Tue, 20 Sep 2022 18:51:33 +0000 (14:51 -0400)]
AMDGPU: Add baseline test for expansion of 16-bit local atomics

The expansion is currently using the wrong pointer size.

2 years agoAtomicExpand: Switch test to generated checks
Matt Arsenault [Tue, 20 Sep 2022 20:03:13 +0000 (16:03 -0400)]
AtomicExpand: Switch test to generated checks

2 years agoVE: Use generated checks for a copy-pasted output test
Matt Arsenault [Tue, 20 Sep 2022 20:32:08 +0000 (16:32 -0400)]
VE: Use generated checks for a copy-pasted output test

2 years ago[clangd] Return earlier when snippet is empty
Tom Praschan [Sun, 18 Sep 2022 16:48:11 +0000 (18:48 +0200)]
[clangd] Return earlier when snippet is empty

Fixes github.com/clangd/clangd/issues/1216

If the Snippet string is empty, Snippet.front() would trigger a crash.
Move the Snippet->empty() check up a few lines to avoid this. Should not
break any existing behavior.

Differential Revision: https://reviews.llvm.org/D134137

2 years ago[HLSL] add sqrt library function
Joshua Batista [Tue, 20 Sep 2022 19:46:06 +0000 (14:46 -0500)]
[HLSL] add sqrt library function

This change exposes the sqrt library function for HLSL scalar types,
excluding long and long long doubles. Sqrt is supported for all scalar, vector,
and matrix types. This patch only adds a subset of scalar type support.

Long and long long double support is missing in this patch because that type
doesn't exist in HLSL.

The full documentation of the HLSL asin function is available here:
https://docs.microsoft.com/en-us/windows/win32/direct3dhlsl/dx-graphics-hlsl-sqrt

Reviewed By: beanz

Differential Revision: https://reviews.llvm.org/D132711

2 years ago[LLDB][NativePDB] Pass register_kind as reference (an oversight in 71d778f).
Zequan Wu [Tue, 20 Sep 2022 20:38:18 +0000 (13:38 -0700)]
[LLDB][NativePDB] Pass register_kind as reference (an oversight in 71d778f).

Fix #57799

2 years ago[NFC][AMDGPU] Refactor AMDGPUDisassembler
Scott Linder [Tue, 20 Sep 2022 20:05:19 +0000 (20:05 +0000)]
[NFC][AMDGPU] Refactor AMDGPUDisassembler

Clean up ahead of a patch to fix bugs in the AMDGPUDisassembler.

Use lit.local.cfg substitutions and more idiomatic use of split-file to
simplify and extend existing kernel-descriptor disassembly tests.

Add a comment to AMDHSAKernelDescriptor.h, as at least one small set
towards keeping all kernel-descriptor sensitive code in sync.

Reviewed By: kzhuravl, arsenm

Differential Revision: https://reviews.llvm.org/D130105

2 years ago[RISCV] Support -mno-implicit-float.
Craig Topper [Tue, 20 Sep 2022 19:13:01 +0000 (12:13 -0700)]
[RISCV] Support -mno-implicit-float.

This can be used to disable vectorization and memcpy/memset
expansion for things like OS kernels. It also disables implicit
uses of scalar FP, but I don't know if we have any of those for
RISC-V.

NOTE: Without this patch you can still do -Xclang -no-implicit-float

Reviewed By: rui.zhang

Differential Revision: https://reviews.llvm.org/D134077

2 years ago[AMDGPU] Fix size of SOPK instructions to 4 bytes
Anshil Gandhi [Tue, 20 Sep 2022 20:23:46 +0000 (14:23 -0600)]
[AMDGPU] Fix size of SOPK instructions to 4 bytes

Instructions in SOPK format may not have 32-bit
literal constants following the instruction.

Differential Revision: https://reviews.llvm.org/D133972

2 years ago[InstCombine] Handle integer extension in `select` patterns using the condition as...
Markus Böck [Tue, 20 Sep 2022 20:12:10 +0000 (22:12 +0200)]
[InstCombine] Handle integer extension in `select` patterns using the condition as value

These patterns were previously only implemented for i1 type but can be extended for any integer type by also handling zext and sext operands.

Differential Revision: https://reviews.llvm.org/D134142

2 years ago[clangd] Remove second tracer (which breaks threading contract)
Sam McCall [Tue, 20 Sep 2022 20:17:02 +0000 (22:17 +0200)]
[clangd] Remove second tracer (which breaks threading contract)

2 years ago[mlir][standalone] Specify python path when configuring
rkayaith [Tue, 20 Sep 2022 02:14:45 +0000 (22:14 -0400)]
[mlir][standalone] Specify python path when configuring

Specifying the python path here ensures that the python binary used matches the
one used by the main MLIR tests. This is useful when cmake's automatic detection
has to be overridden.

Reviewed By: stellaraccident, bondhugula

Differential Revision: https://reviews.llvm.org/D134251